Validating Thermodynamic Models of Arc‐Magma Differentiation and Training Neural Networks for Rapid Thermodynamic Property Inference

Abstract Magmatic systems are the products of the migration and storage of compositionally evolving magmas within the solid crust. Direct computation of thermodynamic properties during magma transport remains a major challenge in multiphase‐transport modeling due to its computational cost. Here, we present a proof‐of‐concept neural network‐based workflow that addresses these challenges. First, we validate thermodynamic databases for modeling arc magma differentiation by comparing the Gibbs Free Energy Minimization (GFEM) software MAGEMin to a compilation of laboratory experiments. Second, we use MAGEMin to generate a data set spanning the entire temperature, pressure, and compositional space relevant to arc magmas. This data set is then used to train a neural network for rapid inference of thermodynamic properties. Results indicate that MAGEMin‐predicted liquid compositions and densities generally agree well with those determined experimentally. The accuracy of mineral phase abundance predictions depends on the starting composition. Notably, the ability to predict amphibole stability by the thermodynamic model is better compared to other thermodynamic packages. The network demonstrates excellent performance and accuracy in predicting the thermodynamic properties of arc magmas. Liquid composition and density are predicted with root‐mean‐square errors below 1 wt.% and 16 kg m −3 , respectively. Network estimates are ca. 3,000 times faster on the CPU compared to GFEM calculations and 10,000 data points can be calculated in under 1.0 ms on the graphics processing unit. Our workflow offers an efficient approach to the computational challenge of integrating composition‐dependent physical properties into reactive transport algorithms. For operational use, the network should be re‐trained on application‐specific data sets.